Location

Brunswick, ME

Start Date

1-1-1990 12:00 AM

Description

Probably the most persistent general problem in acoustic emission (AE) applications is signal source identification. Past applications of pattern recognition techniques to AE have been successful, but require measurement of a parameter such as source location, load, etc., which is well-correlated with specific source types, in addition to AE signal characteristics used as the basis of the feature [1,2]. A training set comprised of a subset of the test data has also been required since good classification features and the distribution of their values are only appropriate when applied to the specific test from which they were obtained. The solution to these problems is to find robust features, or a way to predict features and feature values. Some empirical work along these lines has been done by Pacific Northwest Laboratory (PNL), operated by Battelle Memorial Institute [1,3,4]. In this paper, a transfer function between power spectral density (PSD) feature sets is established to relate the responses of two detection channels to a given source. The method may aid in identifying robust features and in predicting feature value distributions from calibration and a priori information.

Book Title

Review of Progress in Quantitative Nondestructive Evaluation

Volume

9A

Chapter

Chapter 3: Interpretive Signal and Image Processing

Section

A: Signal Processing and Neural Networks

Pages

647-654

DOI

10.1007/978-1-4684-5772-8_81

Language

en

File Format

application/pdf

Share

COinS
 
Jan 1st, 12:00 AM

Transformation of Acoustic Emission Pattern Recognition Features

Brunswick, ME

Probably the most persistent general problem in acoustic emission (AE) applications is signal source identification. Past applications of pattern recognition techniques to AE have been successful, but require measurement of a parameter such as source location, load, etc., which is well-correlated with specific source types, in addition to AE signal characteristics used as the basis of the feature [1,2]. A training set comprised of a subset of the test data has also been required since good classification features and the distribution of their values are only appropriate when applied to the specific test from which they were obtained. The solution to these problems is to find robust features, or a way to predict features and feature values. Some empirical work along these lines has been done by Pacific Northwest Laboratory (PNL), operated by Battelle Memorial Institute [1,3,4]. In this paper, a transfer function between power spectral density (PSD) feature sets is established to relate the responses of two detection channels to a given source. The method may aid in identifying robust features and in predicting feature value distributions from calibration and a priori information.